refactor(models)!: relax the schema-v2 mirrors from extra="forbid" to extra="ignore"

[rahlk]

Stacked on #385 — target it, not release/2.0, so the diff shows only this change. GitHub retargets
automatically when #385 merges.

Taking up an additive analyzer release used to fail the whole payload rather than the one field it
added. codeanalyzer-java 3.1.2 put var on param_in/param_out — backward compatible by
construction — and JAnalysis produced 2515 validation errors on daytrader8 until JParamEdge
declared it.

What this gives up

forbid was the drift detector, and it worked: it surfaced that var within seconds of the pin bump.
Under ignore the same addition is absorbed silently, and the first symptom is a wrong answer from
something reading a field the SDK never learned about — slow and quiet in place of loud and immediate.

The replacement worth building is a comparison against each analyzer's published schema in
codeanalyzer-schema (which codeanalyzer-python#198 / java#252 / typescript#199 now enable): report
what is unmodelled instead of refusing to parse. That decouples parseability from completeness, which
is the actual problem with forbid. Not in this PR.

Scoped narrower than the whole codebase

relaxed	_Base in models/java and models/typescript — the analyzer mirrors, validated from analysis.json — plus JCompilationUnit, which overrides model_config wholesale for its alias settings, so _Base's change never reached it (a subclass config replaces rather than merges)
kept forbid	JCallableOverview / JClassOverview in projections.py

Nothing calls model_validate on the projections — they are constructed by this SDK from graph rows
and never parsed from the wire. Strictness there catches our typo'd keyword argument, not the
analyzer's additions, so relaxing it would surrender a real guard for zero uptake benefit. Say so if
you want it widened; there is a test pinning the distinction either way.

ignore, not allow

allow retains unknown fields in model_extra and so widens model_dump_json() with whatever the
analyzer emitted — and several tests assert properties of dumps (E6's
"can://" not in overview.model_dump_json()). A field the SDK does not model must not be able to
change what a dump contains. Every new test asserts model_extra is empty for that reason.

What this does not cost — now asserted, not assumed

extra does not govern required fields, so "a 1.x analysis.json is refused, not parsed" survives
intact. A v1 payload still fails, because it lacks what v2 requires rather than carrying extras.
test_v1_shaped_payload_is_rejected passes unchanged in both languages, and the new tests say so
beside it, so the two mechanisms are not confused for each other next time this policy is weighed.

The blast radius was found empirically, not by grep

Four tests asserted the old policy and not one contained the word "forbid" — they were named for
the behaviour (test_unknown_field_is_rejected, test_unknown_top_level_key_is_rejected, …), so only
running the full suite located them. Replaced rather than deleted, because the surviving half is worth
writing down: each now asserts an undeclared field is accepted and dropped, so a caller reaching
for unmodelled analyzer data learns it finds nothing rather than an error.

One replacement corrected me rather than the code. I first asserted not hasattr(unit, "file_path")
and got assert not True: file_path is a property over a PrivateAttr that JApplication
stamps from the symbol-table key (models.py:691,714), not a wire field. The test now asserts what
matters — the key is not a model field, the property still resolves from the symbol-table key, and the
injected wire value never took effect. Stronger than the original, which only proved something
raised.

Two comments were actively misleading and are fixed. # extra="forbid": this line is the assertion
annotated a bare _load(level) that now asserts nothing, and the test it sat in claimed that parsing
proved no undeclared field had appeared. That test is renamed and its docstring now states that the
positive assertions carry it, and that a silent pass is the accepted cost of this change rather than
an oversight in the test.

Gate

1599 passed, 370 skipped, no failures — 1597 on #385's branch, +2 net.

Closes #386